DP-preview · client-side parser

AI provenance · playground

Croissant manifest → AI-flavored verdict.

Paste or edit a Croissant 1.1 manifest on the left; the right pane re-parses on every keystroke and renders a verdict against the canonical T-set tag glossary. Three presets, lifted from the AI-provenance example corpus, demonstrate distinct verdict shapes — refutes, suspends, clean. The interactive parser is a client-side TS approximation ofsilver-croissant-v2::parse; the real engine wires up when the WASM bridge ships. See playground.veric.dev for the full set of worked verdicts and /substrate for the substrate-vs-vertical claim.

Preset:

Croissant manifest

input · editable

{
  "_comment": "MOCK fixture — see examples/MOCK-ai.md. Croissant 1.1 shape per https://docs.mlcommons.org/croissant/. Consumed by Agent E's silver-croissant-v2 mock and Agent L's playground.",
  "@context": {
    "@language": "en",
    "@vocab": "https://schema.org/",
    "ml": "http://mlcommons.org/croissant/"
  },
  "@type": "ml:Dataset",
  "name": "fintech-personalization-v3",
  "version": "3.0.1",
  "description": "User-behaviour corpus for fine-tuning the personalization assistant.",
  "license": "internal-use-only",
  "creator": {
    "@type": "Organization",
    "name": "AcmeBank ML Platform"
  },
  "datePublished": "2026-04-22",
  "ml:distribution": [
    {
      "@type": "ml:FileObject",
      "@id": "training-text-parquet",
      "name": "training_text.parquet",
      "contentUrl": "warehouse://feat_training_text",
      "encodingFormat": "application/parquet",
      "ml:source": {
        "warehouse": "snowflake",
        "schema": "ml_features",
        "table": "feat_training_text"
      }
    }
  ],
  "ml:recordSet": [
    {
      "@type": "ml:RecordSet",
      "@id": "training-text",
      "name": "training_text",
      "ml:field": [
        {
          "@type": "ml:Field",
          "name": "user_id",
          "ml:dataType": "ml:Integer",
          "ml:tags": [
            "identifier"
          ]
        },
        {
          "@type": "ml:Field",
          "name": "text",
          "ml:dataType": "ml:Text",
          "ml:tags": [
            "training-input"
          ],
          "description": "Free-text feature concatenated from user metadata + behaviour signals."
        },
        {
          "@type": "ml:Field",
          "name": "label",
          "ml:dataType": "ml:Text",
          "ml:tags": [
            "training-label"
          ]
        }
      ]
    }
  ],
  "ml:provenanceTags": {
    "_comment": "veric-extension. Forbidden-flow rule for this manifest: no `pii`-tagged source may reach any field tagged `training-input`.",
    "forbids": [
      {
        "from": "pii",
        "to": "training-input"
      },
      {
        "from": "personal-data",
        "to": "training-input"
      }
    ]
  }
}

Edits re-parse on every keystroke (client-side).

Verdict

mock · refutes

Manifest summary

name: fintech-personalization-v3
version: 3.0.1
license: internal-use-only
recordSets: 1
distribution: 1

Findings (4)

ERROR · T6_INFORMATION_FLOWml:provenanceTags.forbids[from=pii, to=training-input]
Forbidden-flow precondition met: a field is tagged `training-input`, and the manifest declares `pii → training-input` as a forbidden flow. The full graph proof requires the pipeline IR (M1 mock cannot close it).
ERROR · T6_INFORMATION_FLOWml:provenanceTags.forbids[from=personal-data, to=training-input]
Forbidden-flow precondition met: a field is tagged `training-input`, and the manifest declares `personal-data → training-input` as a forbidden flow. The full graph proof requires the pipeline IR (M1 mock cannot close it).
INFO · GLOSSARY_UNKNOWN
Tag `identifier` is not in the canonical T-set glossary. The verifier will treat it as an opaque marker until added to `content/ai-tag-glossary.json`.
INFO · GLOSSARY_UNKNOWN
Tag `training-label` is not in the canonical T-set glossary. The verifier will treat it as an opaque marker until added to `content/ai-tag-glossary.json`.

Per-field tag analysis

RecordSet	Field	Tags · T-tier · regulations
training_text	user_idml:Integer	identifier · unknown
training_text	textml:Text	T6·training-input·EU AI Act
training_text	labelml:Text	training-label · unknown

What a real verifier does differently

The M1 client-side parser inspects only the manifest. The real verifier (Agent E’s silver-croissant-v2, wired through silver-l8-v2 information-flow) closes the proof against the pipeline IR — it traces tag flow across joins, projections, and emits, and refutes only when a forbidden tag actually reaches a sink. It also discharges per-row contracts (e.g. license-attribution-non-null) against the corpus, not the manifest. This UI swaps to that engine when the WASM bridge ships; the rendering contract stays the same.